Christopher Lemmer Webber is a user on octodon.social. You can follow them or interact with them if you have an account anywhere in the fediverse. If you don't, you can sign up here.

@cwebber expert systems come back from the grave

or were they ever really dead

@KitRedgrave maybe the ai winter was really just an ai cryogenic freeze

@cwebber i would love to know how they deal with ingesting all these papers and generating a suitable knowledge representation. are they using a different machine for that?

@KitRedgrave Oh, it's converting it I think somehow. I'm going to look at what it's using.

@KitRedgrave HAHA AWESOME

if I'm reading this write it loads it all into racket hashtables and then serializes them as racket files github.com/webyrd/mediKanren/b

@KitRedgrave *if I'm reading this right

but oddly appropriate typo

@cwebber so it's a nifty use of minikanren but they have a ways to go

@cwebber @KitRedgrave We had NoSQL. That didn't go far enough. The next step is NoDB. Flat files of serialized data.

@notclacke @KitRedgrave Note that I didn't have the full story:

> There's a little more to it. Greg Rosenblatt created a nice, efficient graph database in Racket that is fast when reading data from disk. Works very well in practice! :)

twitter.com/webyrd/status/1006

@KitRedgrave @cwebber it looks like they use something called SemMedDB

@er1n @cwebber oh cool, somebody went and solved that problem. at least for now

@KitRedgrave @cwebber "The Semantic MEDLINE Database (SemMedDB) [1] is a repository of semantic predications (subject-predicate-object triples) extracted by SemRep, a semantic interpreter of biomedical text [2]."

@er1n Oh I didn't realize @KitRedgrave meant the data... I thought we were talking about the data store!

Not the only cool resource published by the us gov and in the public domain... the USDA nutrition database is also cool as hell ars.usda.gov/northeast-area/be

@cwebber @er1n yeah. reading that many papers and getting the knowledge is not at all trivial! interesting that they had some way of getting that large a knowledgebase... though one has to wonder how valid some of it is given that some branches of science have had bad problems with replicable results

@KitRedgrave @er1n I wonder if it's data the FDA itself used for its evaluation of whether or not a drug may be introduced to the market?

@KitRedgrave @er1n @cwebber No, as it hasn't been nearly validated enough. That might help guide what to study, but you are never going to beat a properly done three-phase trial of thousands of people for finding problems with a drug.

@KitRedgrave @cwebber

they hibernate periodically, waiting for strong enough computers to give them the ability to dominate the planet
@KitRedgrave @cwebber this thing isn't Dr. House unless it insults patients and almost kills them before coming up with the right diagnosis

though it is AI, so it'll probably get the killing patients bit right
@cwebber @KitRedgrave I wouldn't mind an AI that could do Hugh Laurie's American account, though
@KitRedgrave @cwebber

> In 15 seconds, the program has an answer: Gleevec, a brand name for imatinib, was designed to treat cancer. But it also turns out to inhibit the BCR gene, which causes asthma. Why does mediKanren think so? The program serves up links to the relevant papers on PubMed, the NIH’s vast research database.

I love that this expert system even explains itself and offers supporting evidence. That's some next-level expert system stuff.